Export and endpoints

About

There are two directions for trace data: out of Observe (exporting what you’re viewing) and into Observe (the OTLP endpoints traceAI sends to). This page covers both.

Export from the trace explorer

The download icon in the trace explorer header exports the current view — the traces that match your active filters and time range.

CSV is the only export format. There is no JSON or Parquet download from the trace explorer.

Format	Use for
CSV	Spreadsheet analysis, sharing, importing elsewhere.

The CSV holds one row per span in the current view, with the columns the explorer shows: trace ID, span name, status, model, provider, latency, token counts, cost, timestamp, and any eval scores present on the span.

Ingestion endpoints

traceAI exports spans over OTLP to FutureAGI. The transport and target are environment-driven:

Variable	Transport	Default
FI_BASE_URL	HTTP collector	FutureAGI cloud collector
FI_GRPC_URL	gRPC collector	FutureAGI cloud collector

When unset, both variables default to the FutureAGI cloud collector, so a cloud project needs no endpoint configuration — only FI_API_KEY and FI_SECRET_KEY.

• Cloud: leave the defaults; set FI_API_KEY and FI_SECRET_KEY.
• Self-hosted: point FI_BASE_URL / FI_GRPC_URL at your own collector host so spans stay in your network.

Choose the transport with transport=Transport.HTTP (default) or Transport.GRPC in register(). See Set up tracing.

Endpoint contract

The ingestion endpoint is an OTLP/traces receiver — you don’t call it directly; the traceAI SDK’s exporter does. The contract:

Aspect	Detail
Protocol	OTLP over HTTP (protobuf) or gRPC — the OpenTelemetry standard, not a proprietary API.
Operation	Export spans (write-only). There is no read/query endpoint; reading happens in the trace explorer.
Auth	FI_API_KEY + FI_SECRET_KEY from the keys page, sent by the SDK on every export. Keys are workspace-scoped.
Success	The exporter batches spans and sends them in the background; a successful export returns no payload. Spans appear in Observe within seconds.
Errors	A 401 means the keys are wrong for this workspace; a 4xx means a malformed/oversized batch; transient 5xx/network errors are retried by the batch exporter.
Limits	Spans are sent by the batch span processor on an interval, so a short-lived process must call trace_provider.force_flush() before exit or the last batch is lost. Very large payloads (huge prompts/outputs) can be dropped: cap attribute size or mask them at the SDK. Future AGI Cloud accepts OTLP requests up to 16 MiB over HTTP or gRPC. See Ingestion request limits.
Versioning	Pin fi-instrumentation-otel and each instrumentor to a tested version so a release can’t change span shape under you; the wire format follows the OTLP version the SDK ships.

Ingestion request limits

Future AGI Cloud enforces the same supported request limit for both OTLP transports:

Transport	Maximum request size	Error when exceeded
OTLP/HTTP	16 MiB	413 Request Entity Too Large
OTLP/gRPC	16 MiB	RESOURCE_EXHAUSTED

An oversized request is rejected as a whole and its spans are not stored. If you see either error in your exporter logs, reduce the number of spans sent in each batch. OpenTelemetry’s default maximum export batch size is 512 spans. Start with:

OTEL_BSP_MAX_EXPORT_BATCH_SIZE=256

Related